A guide to big data workload- management challenges
A guide to big data workloadmanagement challenges
By George Gilbert
This research was underwritten by DataStax.
BIG DATA
Table of contents
EXECUTIVE SUMMARY
4
UNDERSTANDING THE NEW CLASS OF APPLICATIONS
5
New applications supporting new business models
5
Requirements for new technology underpinnings
6
VOLUME, VELOCITY AND VARIETY OF DATA
6
Data volume
7
Data velocity
8
Data variety
8
REAL-TIME, MASSIVELY SCALABLE AND CLOSED-LOOP
CHARACTERISTICS OF APPLICATIONS
10
Real time
10
Massively scalable
11
Closing the loop
12
APPROACHES TO SCALABILITY
14
Traditional SQL DBMS
14
Emerging NoSQL databases
15
Cassandra
15
ALTERNATIVES TO CASSANDRA
16
Amazon DynamoDB
16
Oracle NoSQL
16
HBase
17
Sustainable versus nonsustainable differentiators
17
DRIVING BETTER DECISIONS INTO ONLINE APPLICATIONS
18
A note on the distinction between Cassandra and DataStax Enterprise
A guide to big data workloadmanagement challenges
May 2012
21
-2 -
BIG DATA
KEY TAKEAWAYS
23
Business considerations
23
Technology considerations
23
APPENDIX A: UNDERSTANDING THE ASSUMPTIONS IN NOSQL
DATABASES AS A CLASS OF SYSTEMS RELATIVE TO SQL DATABASES 25
ABOUT GEORGE GILBERT
26
ABOUT GIGAOM PRO
26
FURTHER READING
27
A guide to big data workloadmanagement challenges
May 2012
-3 -
BIG DATA
Executive summary
The explosive growth in the volume, velocity, variety and complexity of data has
challenged both traditional enterprise application vendors as well as companies built
around online applications. In response, new applications have emerged, ones that are
real-time, massively scalable and have closed-loop analytics. Needless to say, these
applications require very different technology underpinnings than what came before.
Traditional applications had a common platform that captured business transactions.
The software pipeline extracted, cleansed and loaded the information into a data
warehouse. The data warehouse reorganized the data primarily to answer questions
that were known in advance. Tying the answers back into better decisions in the form
of transactions was mostly an offline, human activity.
The emerging class of applications requires new functionality that closes the loop
between incoming transactions and the analytics that drive action on those
transactions. Closing the loop between decisions and actions can take two forms:
Analytics can run directly on the transactional database in real time or in closely
integrated but offline tasks running on Hadoop. Hadoop typically supports data
scientists who take in data that’s far more raw and unrefined than the type found in a
traditional enterprise data warehouse. The raw data makes it easier to find new
patterns that define new analytic rules to insert back into the online database.
This paper is targeted at technology-aware business executives, IT generalists and
those who recognize that many emerging applications need new data-management
foundations. The paper surveys this class of applications and its technology
underpinnings relative to more-traditional offerings from several high-level
perspectives: the characteristics of the data, the distinctiveness of the new class of
applications, and the emerging database technologies — labeled NoSQL, for lack of a
better term — that support them. Although the NoSQL label has been applied to many
databases, this paper will focus on the class of systems with a rich data model typified
A guide to big data workloadmanagement challenges
May 2012
-4 -
BIG DATA
by Cassandra. Other databases in this class include HBase, DynamoDB and Oracle
NoSQL.
Understanding the new class of applications
Consider large-scale machine-to-machine monitoring and measurement as an
example of applications that require new technology underpinnings. E3 Greentech, a
Decatur, Ga.–based startup, has a great example of such an application. Although
currently targeted at utilities and others that generate, distribute and consume power,
it can ultimately apply to any intelligent grid of machines.
Power is now generated not only by utilities at large, central sites but also at
distributed sites with basic generators. The latter can also store power and distribute it
back into the grid when necessary. More important, intelligent sensors and devices are
making it possible for homes to instrument their electric appliances — from water
heaters and air conditioners to dishwashers and entertainment centers — in order to
measure and manage electricity consumption.
New applications supporting new business models
One notable aspect of E3 Greentech’s application, and potentially similar ones, is how
it can enable new business models. E3 Greentech’s application makes it possible for
utilities to not only provide electricity but also offer home energy management
services. Because the application is part of an ongoing relationship between a utility
and its end consumers, E3 Greentech can support pricing that captures some of this
ongoing value. These pricing models reflect a managed service that goes beyond
typical SaaS application pricing, which only captures the value of managing the
application itself.
A guide to big data workloadmanagement challenges
May 2012
-5 -
BIG DATA
Requirements for new technology underpinnings
The first requirement for this class of application is that the same database has to serve
both as the operational data store as well as the basis for real-time analytics. In
addition, analytics that are performed offline ideally should be on another cluster of
the same database so that moving data back and forth doesn’t require the complicated
and slow pipeline that typically connects traditional online transaction processing
(OLTP) databases and enterprise data warehouses. The database must also deal with
data that is far less refined than in traditional SQL databases. It must include not only
structured data but also semistructured and unstructured data. A related requirement
is the ability to handle potential changes in the format of the data without taking the
database or application offline. Finally, the database needs to run on premise, in the
cloud or on a combination of both without complicated configuration of the data
distribution mechanics.
Volume, velocity and variety of data
The previously nebulous definition of “big data” is growing more concrete as it
becomes the focus of more applications. As seen in Figure 1 (below), volume, velocity
and variety make up three key characteristics of big data:
! Volume. Rather than just capturing business transactions and moving samples
and aggregates to another database for analysis, applications now capture all
possible data for analysis.
! Velocity. Traditional transaction-processing applications might have captured
transactions in real time from end users, but newer applications are increasingly
capturing data streaming in from other systems or even sensors. Traditional
applications also move their data to an enterprise data warehouse through a
deliberate and careful process that generally focuses on historical analysis.
! Variety. The variety of data is much richer now, because data no longer comes
A guide to big data workloadmanagement challenges
May 2012
-6 -
BIG DATA
solely from business transactions. It often comes from machines, sensors and
unrefined sources, making it much more complex to manage.
Figure 1: The three characteristics of big data: volume, velocity and variety
Source: IBM
!
Data volume
A more detailed look at data volumes highlights some of the new database
requirements. E3 Greentech’s application takes in far more data than simply reading a
home’s meter every 30 minutes to find out aggregate power consumption. In fact, the
application reads the consumption data on so many sensors tied to so many appliances
and devices that it takes readings roughly 100,000 times per day per residence. As a
result, E3 Greentech needs a database that can handle more writes than a traditional
database running on one big server or on one shared storage device.
A guide to big data workloadmanagement challenges
May 2012
-7 -
BIG DATA
Data velocity
Measuring and managing energy consumption down to the device level requires realtime data and analysis. To look at E3 Greentech again, that means collapsing the
traditional pipeline that batch processes data between one database that handles
writes and another that handles reads for analysis. E3 Greentech handles data more
like a nonstop incoming stream with a set of continuous queries spitting out answers.
The company has already tested Cassandra at 1 million writes per second on a cluster
of servers. And that doesn’t count the simultaneous queries that are analyzing how to
manage the demand for electricity for maximum efficiency. That still leaves plenty of
headroom.
Part of what makes this speed possible is that the incoming data is primarily new
inserts, not updates to existing data, which requires more overhead. In addition, the
real-time queries tend to be less complex than what is possible with a traditional SQL
database.
To be fair, there are queries in E3 Greentech’s application, and with most applications
in this class, that require more processing time than a real-time system can handle.
Here Cassandra’s Hadoop integration becomes relevant. Another cluster holds a
replica of the Cassandra data in Hadoop for offline queries while the answers get fed
back into the real-time Cassandra cluster to improve its energy-management
decisions.
Data variety
The variety of data that the new class of databases has to manage is arguably their
most fundamental differentiator.
SQL databases derive much of their strength from the order and discipline they force
on their data. In the example of an energy-management application, every device or
A guide to big data workloadmanagement challenges
May 2012
-8 -
BIG DATA
entity being measured and monitored would have to be accounted for and described
down to its very last attribute. This schema structure requires the application
designers to go through a very detailed process of refining the data until everything is
classified. The value of this approach is that there is no ambiguity in the data, and the
application knows exactly what it’s dealing with. The downside is that if the application
needs to handle new devices or query the data to find new signals not previously
anticipated in the noise of all the data, a lot of additional refining has to happen.
Cassandra makes it much easier to avoid the upfront refining process, as long as the
application is intelligent enough to make sense of the updated or new device entities. It
also makes it easier to ask many new questions of the data without pumping it through
the batch process to get it into the data warehouse (though more-sophisticated
questions would still have to go to the Hadoop cluster). In terms of new devices, if the
device readings from a house started to include a new type of outlet that also included
an occupancy sensor and a temperature sensor, the same column family would start
storing the additional attributes without missing a step. Meanwhile the application has
already been written to understand that it can now track temperature and occupancy
readings from this new device.
From another perspective, it’s critical to understand that all the raw data, including
these new readings, is available for analysis. The traditional batch process of
extracting, transforming and loading data from the online database to the data
warehouse removes any variation in the type of data. Again, there is a purpose in this
type of refinement. It makes it easier for analysts to make sense of the data, although
at the expense of being able to go back into the messy details to find new signal within
the noise.
A guide to big data workloadmanagement challenges
May 2012
-9 -
BIG DATA
Real-time, massively scalable and closed-loop
characteristics of applications
Real time
Continuing with the home energy management example, reducing consumption
demands much more than just switching off everyone’s air conditioner when the
weather is hot and humid. Managing real-time consumption minimizes electricity
charges during peak usage, and it can meaningfully reduce the 30 to 40 percent of
power-generating capacity that is built just to satisfy peak demand. The application
depends on the real-time read/write database to “turn the crank” and evaluate the
state of all the assets under management in a few milliseconds as new readings
continue to stream in.
That’s where real-time applications and databases come into play. The application can
compare the trend of the recent readings of indoor and outdoor temperatures and the
current thermostat settings for all houses in a certain area that share the same
weather. Even though these readings are streaming in, in real time, the database can
store the data so it can be accessed instantly and without blocking additional incoming
writes. Furthermore, offline analysis from the application running on the replica
cluster that runs Hadoop might calculate and periodically update a thermal envelope
estimate for each house. By combining these readings plus an occupancy sensor, the
application can determine when to raise or lower the thermostat.
For such a system, Cassandra stores and organizes not only the readings but also the
components that control the application’s mechanics. The application has to put
together groups of entities — devices, sensors, homes, neighborhoods and weather
regions — dynamically. And as described in the section on data variety, it has to make
sense of these groups even when the data coming back from the individual readings
changes or as new devices are added. In order to make this happen, Cassandra stores
each device and those that are similar in a group of rows called column families. The
A guide to big data workloadmanagement challenges
May 2012
- 10-
BIG DATA
values for each sensor reading or row get stored in a flexible number of columns. If the
device gets swapped for a similar type, it can store additional or fewer columns for
each reading.
The magic is how the system can calculate in real time groups of geographic regions,
homes and devices that meet changing criteria and on which the application is going to
take action to change energy consumption. The incoming stream of sensor data might
not conform exactly to what’s stored in the database, but the application still correlates
the individual readings from the physical devices with the dynamic collections of
classes of devices organized in the database.
Massively scalable
In addition to real-time responsiveness, it takes an application running on a massively
scalable database like Cassandra to manage the energy consumption of potentially
millions of individual devices and appliances across hundreds of thousands of homes
relative to the peak load on the grid. The scalability comes from being able to apply the
correct power-management rules for each dynamic group described above (i.e., turn
off the water heater for homes when the temperature is above 80 degrees and no one is
home).
There are an unlimited number of possible rule combinations as they are applied to
progressively more-targeted groupings within the hundreds of thousands of individual
houses under management. For example, the application might get a reading that a
particular home just became occupied. It then runs a rule to reset the thermostat back
to where the occupant originally left it if the temperature outside is more than 10
degrees higher than inside. Once that rule executes, the application might use another
to evaluate the same temperature readings to determine just how much additional
power the water heater needs to reach a comfortable level. Operating on this scale
A guide to big data workloadmanagement challenges
May 2012
- 11-
BIG DATA
while handling reads and writes requires a database that can run on a large cluster of
machines without bottlenecking on data stored and shared in a single place.
Most traditional SQL databases for OLTP operate on data that is shared on a common
storage array. They are configured this way so that any node in the database cluster
can get at any data item or set of items. And for those that do scale out to a cluster of
independent machines without shared storage, they still have the requirement to
separate the database handling the high volume of updates from the one configured to
ask the mostly predefined set of queries. The pipeline that connects these two
databases via a batch process is the biggest impediment to real-time operation.
Closing the loop
In traditional applications, analytics haven’t been part of a continuous loop that
automatically feeds better decisions back into real-time transactions. With Cassandra,
there are two mechanisms to accomplish this. The real-time section above describes
how the E3 Greentech application dynamically groups devices and homes in order to
apply rules that improve energy efficiency. This part of the analytics works in real time
because there is no batch-processing pipeline between the online application and the
analytics. But there are also patterns of analysis that need access to more data and
have to run offline in order to deliver meaningful answers. Cassandra’s deep
integration with Hadoop supports these scenarios and makes it easy to feed the
answers back into the real-time part of the application.
While the real-time part of the application may be operating on current sensor
readings, Hadoop works best for finding patterns and correlations across large data
sets. For example, a smart-grid application needs to know the average daily
temperature for each geography as well as a distribution of readings over time. With
this information, the Hadoop-sourced analysis can be fed back periodically into the
A guide to big data workloadmanagement challenges
May 2012
- 12-
BIG DATA
real-time application loop. There it can help make decisions based on where the
current temperature is relative to its average as well as the highs and lows.
This same type of pattern analysis can be useful for supporting offline decisions. The
application might measure how quickly the temperature changes in all homes in a
neighborhood relative to weather conditions. Based on the results, it can recommend
to the property manager which ones can benefit from weatherization. Alternatively, it
can look at the efficiency relative to manufacturer benchmarks of heavy-duty
appliances like the water heaters and the air conditioners within each home. After
recommending what can be fixed or upgraded, the application can again measure their
new performance against the benchmark.
The integration between Cassandra and Hadoop dramatically simplifies these
processes. Cassandra replicates its data with a cluster running the Hadoop Distributed
File System (HDFS), and that runs the MapReduce application framework on top of it.
Because it is a replica cluster, the data gets updated from the online application in real
time. And because it is HDFS, any application written to MapReduce or one of the
higher-level programming frameworks works without modification.
As documented comprehensively elsewhere, MapReduce simplifies the process of
working with massively unrefined data sets such as those from sensors. Rather than
going through the rigorous process of modeling each potential device on the smart grid
and all of its potential attributes, a MapReduce job can analyze just the available
attributes. However, the repository is unrefined and contains all the “noise,” in terms
of additional stray attributes and new sensors. Having that unrefined data already as
part of the repository makes it easier to update the analysis in order to search for new
patterns. The parallelism built into MapReduce makes the process highly scalable.
In traditional enterprise data warehouses, the batch process pipeline connecting from
the OLTP database needs to be updated to handle additional entities in the form of
A guide to big data workloadmanagement challenges
May 2012
- 13-
BIG DATA
new devices or attributes. The data warehouse would also need updating in two
dimensions. First, it would need to accommodate the new information in its refined
format. And it would need to accommodate the new questions developers want to ask.
It should be noted, however, that Hadoop’s flexibility has a price. As described earlier,
unrefined data requires more effort on the part of the developer to make sense of it
and analyze it.
Approaches to scalability
Traditional SQL DBMS
Traditional SQL databases must trade unlimited scalability in favor of maintaining
data consistency and integrity above all else. Managing a bank account is a perfect
example. If a customer withdraws cash, the database has to make sure the transaction
updates the customer’s account balance before any other account activity can take
place. Otherwise it could be possible for multiple withdrawals to exceed the account
balance. That block on any simultaneous activity is the core of the data integrity
features of traditional SQL databases. But that block is also the bottleneck that limits
the ultimate scalability of online-transaction-processing applications on SQL
databases. All processes in an application ultimately must share access to a single
master copy of data in order to ensure there are no reads of “dirty” or out-of-date data.
A clarification on SQL DBMS technical nomenclature:
The full technical requirement with traditional SQL DBMS is for ACID transactions,
which means the database must enforce atomicity, consistency, isolation and
durability. In this context, consistency refers to data-integrity constraints. These
constraints might enforce rules such as requiring all phone numbers to include 10
numeric digits or that a customer record can’t be deleted if it still has outstanding
orders. This definition of consistency is different from how it is applied with
emerging NoSQL databases, as explained in the following section.
A guide to big data workloadmanagement challenges
May 2012
- 14-
BIG DATA
Emerging NoSQL databases
NoSQL databases generally relax the consistency described above in favor of scalability
or availability. The relevant technical acronym is CAP, which stands for consistency,
availability and partition tolerance. These attributes represent the choice of trade-offs
that different NoSQL databases make. Theoretically it’s only possible to offer two of
these three features at any one time in a distributed database.
Because the data doesn’t have to be 100 percent consistent all the time, applications
can scale out to a much greater extent. By relaxing the consistency requirement, for
example, NoSQL databases can have multiple copies of the same data spread across
many servers or partitions in many locations. The data is instead eventually consistent
when the servers are able to communicate with one another and catch up on any
updates one may have missed.
Cassandra
Cassandra lets the application developer dial in the appropriate level of consistency
versus scalability or availability for each transaction. This “tunable consistency” is a
level of sophistication that other databases such as HBase don’t always offer. However,
the extra sophistication comes with more of a learning curve for the developer.
Cassandra also has an in-memory cache to speed access to the most important data on
designated servers in the cluster. However, its current implementation largely
leverages the functionality already in the operating system. By contrast, Couchbase, for
example, includes a cache that intelligently moves data in and out of memory on each
server and across partitions residing on other servers as needed. Despite this feature,
Couchbase is not a substitute for Cassandra. Cassandra offers a richer data model that
combines elements of tabular data in SQL databases and the flexibility typical of rich
key-value stores. E3 Greentech, for example, uses Couchbase to cache and store the
A guide to big data workloadmanagement challenges
May 2012
- 15-
BIG DATA
most recent sensor reads for each device, but it runs the rest of the application on
Cassandra.
Alternatives to Cassandra
Amazon DynamoDB
Amazon recently introduced a new database, DynamoDB, which offers more
sophistication and scale than its earlier, simple key-value store, SimpleDB. Three
attributes distinguished it from competitors when it was introduced:
•
It is a managed service. Customers don’t have to know anything about
deploying or managing a database cluster.
•
It is both configured and priced by a combination of “tunable” read/write
throughput and total storage capacity, so the service itself figures out how to
configure the data.
•
It is designed exclusively to exploit the high throughput and durability of
flash SSD storage. The latest release of Cassandra can also be configured to
support SSD.
Oracle NoSQL
Oracle built its NoSQL database on the foundation of Berkeley DB, which the company
acquired six years ago. The technology shares some similarities with Cassandra in that
it is a distributed, replicated key-value store that runs on a cluster of machines.
However, Oracle has made some design decisions that point to different use cases.
Rather than accommodating arbitrarily large-scale deployments, the database appears
best designed to be deployed in a single rack or an Oracle-designed hardware
appliance. Because it is not a fully peer-t0-peer system, if a master node goes down
and it is configured for relaxed consistency, updates could get lost before they find
A guide to big data workloadmanagement challenges
May 2012
- 16-
BIG DATA
their way to a new master node elsewhere on a network when it gets elected. That type
of scenario wouldn’t happen if it were deployed in a single rack or an appliance.
HBase
HBase is the no NoSQL distributed database that forms part of the Hadoop stack. It
provides strongly consistent reads and writes versus an “eventually consistent” data
store. It was created at Google and derived from a product called BigTable and Google
File System. Cassandra was created at Facebook and while implementing the BigTable
data model; it uses a system inspired by Amazon’s Dynamo for storing data. (Much of
the initial development work on Cassandra was performed by two Dynamo engineers
recruited to Facebook from Amazon.) These differing histories have resulted in
HBase’s being more suitable for data warehousing and large-scale data processing and
analysis (i.e., indexing the Web) and Cassandra’s being more suitable for real-time
transaction processing and serving interactive data.
Sustainable versus nonsustainable differentiators
Some products, such as DynamoDB, have valuable differentiators, but they aren’t
necessarily permanent. For example, a PaaS vendor such as Heroku could offer
Cassandra or HBase as a managed service. Also, in its most recent release Cassandra
has been modified to allow certain workloads to target flash SSD as its storage medium
along with others targeting traditional spinning disk drives. With that modification,
any database can be configured to deliver a promised level of throughput. A managed
service provider could then be simply configured and priced based on throughput, just
like Amazon’s DynamoDB. A customer running it on premise could also achieve much
higher and more predictable performance than running it with spinning disk drives.
That Cassandra can offer configurable support for flexibly distributing data across
multiple on-premise data centers and public clouds is a critical choice. This flexibility
can solve this bottleneck when latency and bandwidth limitations prevent large data
A guide to big data workloadmanagement challenges
May 2012
- 17-
BIG DATA
sets — in the billions of rows — from being moved to or from cloud providers like
Amazon.
Pricing options are also valuable. Having an option to use different storage media is
important also for pricing flexibility. Customers pay dearly for the privilege of running
on flash SSD. Once there, using DynamoDB for a 1 TB database with a throughput of
500 writes per second would cost roughly $70,000 per month.
Driving better decisions into online applications
Making analytics an integral part of online applications requires progressively iterating
on a model to improve its predictive capability. As the model improves, it becomes
easier to benchmark activity captured in the application. Benchmarking makes for
more-effective predictions of outcomes. It also makes detecting anomalies more
accurate so the application can take remediating actions.
As seen in Figure 2, traditional SQL DBMS refine the data they store so that it is easy
for application developers to take advantage of it. However, the refining process
creates a requirement to collect just the data relevant for the ultimate questions it will
answer.
A guide to big data workloadmanagement challenges
May 2012
- 18-
BIG DATA
Figure 2: The data pipeline
Source: Dave Campbell keynote, VLDB 2011
NoSQL databases such as Cassandra don’t necessarily replace SQL databases in
building analytics into online applications. But the new class of databases definitely
enhances the process through complementary functionality in the online database as
well as in the Hadoop cluster.
It’s worth calling special attention to how this process complements analytics in
traditional enterprise data warehouses. Virtually all traditional data warehouses are
based on a combination of aggregated data and samples of the full details in the raw
data set. That is a natural part of the process of modeling or refining the data to fit into
the transactional application and then to batch process it into the data warehouse.
The fundamental trade-off at the heart of this design is to feature data consistency and
integrity at the expense of unanticipated flexibility in the data. Big data supported by
NoSQL data stores such as Cassandra makes the opposite trade-off. It makes it easy to
capture all the detail data in all of its partially structured or unstructured clutter.
Getting all of that unrefined data clutter requires capturing it further back in the data
pipeline than the refined data feeding traditional OLTP applications or data
A guide to big data workloadmanagement challenges
May 2012
- 19-
BIG DATA
warehouses. Access to that raw data makes it possible to pursue a certain line of
analysis that wasn’t previously modeled into the traditional data. The raw data also
improves the predictive models solely through the availability of a bigger data set
against which they are tested. The smart-grid application provides a perfect example.
Figure 3: Systems based on big data
Source: Loosely based on Dave Campbell’s keynote, VLDB 2011
The smart-grid application is an early example of the Internet of things, where every
electrical device in a home is instrumented to report on its operation. To start out, the
production analytic model might measure how much the temperature in each
unoccupied house rises when the A/C is off relative to the outside temperature. Later,
an analyst might use the Hadoop cluster to factor in additional variables such as
humidity and the capacity of the A/C unit and then derive how quickly a particular
house can be cooled. Although the rule comes out of an offline, batch process in
Hadoop, it would get fed back into the online application cluster. Every time a new
reading comes streaming in, the devices, houses, neighborhoods, weather and climate
get evaluated to see if the new rule should run. The time to cool a house to the initial
A guide to big data workloadmanagement challenges
May 2012
- 20-
BIG DATA
thermostat setting becomes one of the thousands of rules running every few
milliseconds as new data from across the system arrives.
The key Cassandra features reviewed here make it an ideal mixed-use database that
can support real-time transactions as well as real-time and batch analytics.
A note on the distinction between Cassandra and
DataStax Enterprise
DataStax is the company behind the Apache project, employing roughly 90 percent of
the committers on the open-source project. The application example in this paper, E3
Greentech, currently uses the open-source version of Cassandra primarily because the
company was started before DataStax Enterprise (DSE) was available. This section
details the features that could have been utilized if the commercial version had been
available in time.
DSE extends Cassandra to simplify deployment and operation and more tightly
integrate Hadoop-based analytics. The free, open-source version of Cassandra has
real-time, two-way replication with HDFS-based Hadoop clusters, but E3 reports it is
extremely complicated to configure. The DSE version, by contrast, works out of the
box. An administrator simply distinguishes the online nodes and the Hadoop analytic
nodes.
The Hadoop nodes are themselves part of DataStax Enterprise, having added several
critical new features. DSE replaces the open-source HDFS file system and introduces
instead the HDFS-compatible Cassandra File System (CFS). Until recently CFS
eliminated the single point of failure in HDFS. If one central node went down, the
Hadoop cluster would be unavailable. HDFS has since fixed this problem. However,
having an HDFS-compatible CFS supports MapReduce and other elements of the
expanding Hadoop stack without modification.
A guide to big data workloadmanagement challenges
May 2012
- 21-
BIG DATA
DSE and its integral version of Hadoop can be deployed with a more configurable
topology that enables greater availability or reduced latency. A data center, which is
DataStax’s name for a cluster, can be configured to have one or more replicas of the
online database as well as Hadoop colocated. In addition, other locations themselves
can have a configurable number of real-time replicas.
Finally, DSE has a tool called OpsCenter for managing database clusters that’s not in
the open-source version. It provides detailed, visual configuration information on the
whole cluster in one place. It relieves the administrator of having to configure and
manage each node through an API.
For more on the innovation driving Hadoop and related technologies, see the
companies driving the main distributions, Cloudera and Hortonworks. They are
driving a very rapid expansion in the complementary developer and administrator
tools and frameworks that make Hadoop accessible to a much broader audience. Many
ISVs and corporate developers have little experience in programming and
administering highly distributed systems.
A guide to big data workloadmanagement challenges
May 2012
- 22-
BIG DATA
Key takeaways
Business considerations
Business managers collaborating with IT have to assess a new class of technologies
that can support new business models:
! Traditional enterprise applications have focused on delivering operating
efficiencies, principally in the general and administrative line item of profit and
loss (P&L) financial statements.
! More recently familiar consumer websites have pioneered businesses where the
online service is itself the business and drives the entire P&L. These same sites
have also pioneered a new class of foundation technologies that are more real
time, more scalable and more closely integrated with online decision making
than traditional applications.
! These new foundation technologies are becoming mature enough for moremainstream deployment. Successful deployment no longer requires the ability to
submit patches to the community managing the source code.
! Businesses that have ongoing relationships with their customers should evaluate
whether they can deliver new services that complement their existing activities
and whether these new foundation technologies can make that possible.
Technology considerations
! Despite the hype cycle around big data and NoSQL databases, they underpin a
major new class of applications.
! Real-time, massively scalable applications with closed-loop analytics differ from
traditional enterprise applications and webscale applications built on traditional
databases in several key ways:
1. In the characteristics of the data they manage
A guide to big data workloadmanagement challenges
May 2012
- 23-
BIG DATA
2. In the combination of real-time and offline closed-loop analytics
they require
3. In the need to combine highly scalable writes and analytics in the
same database so the application can effect improved operational
performance on the real-world systems that it manages itself or via
human operators
4. In the way they can run, easily manage and reintegrate offline
analytics on a rich set of data types on massively scalable
infrastructure
! The database underpinnings of this class of applications have some distinctive
characteristics. In particular, in order to support a high volume of simultaneous
writes and analysis, these databases put a greater burden on developers to keep
track of how their data is stored.
! Relevant databases in this family include Cassandra, Amazon DynamoDB and
Oracle NoSQL.
A guide to big data workloadmanagement challenges
May 2012
- 24-
BIG DATA
Appendix A: understanding the assumptions in
NoSQL databases as a class of systems relative to
SQL databases
The debate about how to manage big data revisits arguments about databases
previously thought settled for decades. For nearly 25 years relational databases have
provided the foundation for enterprise and online applications. What distinguished
this foundation from previous database technologies was the careful, modular
separation between the data and the applications built on top of it. That modular
separation made it possible for applications to worry only about what data they
needed, not how to get at it. Each could be changed and updated without breaking the
other.
Now the big data meme is reopening the debate about how best to manage data. In
some cases, the modular separation of applications and databases will have to give
way. For lack of a better term, NoSQL has become the label for many of the new
approaches.
Within all the new approaches, including the ones that attempt to stretch the
capabilities of RDBMS, there is a common theme: The formal separation that permits
applications to avoid having to worry how the data is stored disappears. While that
sounds like a major step backward to pre-RDBMS days, in some scenarios such as
extreme scalability or managing “messy” data, it makes sense. When performance or
functional requirements demand capabilities unavailable in off-the-shelf modules, the
application developer is responsible for adding those capabilities through proprietary
integration with the data-management layer. Applications have to be more tightly
fused with their data. Once again, they have to know how to get at their data physically
rather than just say what they need. This trade-off between modularity and integration
should be familiar to all readers of Clayton Christensen’s books on disruptive
innovation.
A guide to big data workloadmanagement challenges
May 2012
- 25-
BIG DATA
About George Gilbert
George Gilbert is the co-founder of TechAlpha Partners, a management consulting and
research firm that focuses on the business and technical implications of the transition
to cloud-based data services. Previously Gilbert was the lead enterprise software
analyst for Credit Suisse First Boston, one of the leading investment banks to the
technology sector. Prior to that he worked at Microsoft as a product manager and
spent seven years in product management and marketing at Lotus Development. He
holds a BA in economics from Harvard University.
About GigaOM Pro
GigaOM Pro gives you insider access to expert industry insights on emerging markets.
Focused on delivering highly relevant and timely research to the people who need it
most, our analysis, reports and original research come from the most respected voices
in the industry. Whether you’re beginning to learn about a new market or are an
industry insider, GigaOM Pro addresses the need for relevant, illuminating insights
into the industry’s most dynamic markets.
Visit us at: http://pro.gigaom.com
A guide to big data workloadmanagement challenges
May 2012
- 26-
BIG DATA
Further reading
Locating data centers in an energy-constrained world
Globally data centers are currently buying $30 billion of power per year, but changes
in the power market are rippling through the Internet industry, altering both the
location of data centers and their sources of power. This report parses the many
complications of picking a data center location, which we call data-center siting. Siting
includes considerations beyond the geographic location, such as how to procure energy
and green-energy models, all to run better on a new, ever-shifting energy
infrastructure. This report will help decision makers like CIOs and data-center
planners weigh the most important parameters behind locating new data centers,
whether in the U.S. or globally. Doing so will lead to more cost-effective, reliable and
sustainable data centers.
Sector RoadMap: Hadoop platforms 2012
For years, technologists have been promising software that will make it easier and
cheaper to analyze vast amounts of data in order to revolutionize business. More than
one solution exists, but today Hadoop is fast becoming the most talked about name in
enterprises. There are now more than half a dozen commercial Hadoop distributions
in the market, and almost every enterprise with big data challenges is tinkering with
the Apache Foundation–licensed software. This report examines the key disruptive
trends shaping the Hadoop platform market, from integration with legacy systems to
ensuring data security, and where companies like Cloudera, IBM, Hortonworks and
others will position themselves to gain share and increase revenue.
How Amazon’s DynamoDB is rattling the big data and cloud markets
The latest AWS offering to rock the technology establishment is DynamoDB, a NoSQL
database service that puts the power of NoSQL in the hands of every developer. This
research note analyzes the multiple ways in which Amazon’s announcement has
disrupted the big data and cloud computing market and what that means for other
A guide to big data workloadmanagement challenges
May 2012
- 27-
BIG DATA
companies and offerings in the space — from the startups selling Hadoop distributions
to public cloud providers like Rackspace and Microsoft, which will have to scramble to
keep up and differentiate.
A guide to big data workloadmanagement challenges
May 2012
- 28-
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement